DIRA Workshop and Challenge
Diagram Image Retrieval and Analysis (DIRA):
Representation, Learning, and Similarity Metrics
First workshop on diagram image retrieval and analysis in computer vision
To be held in conjunction with CVPR 2020 in Seattle, WA, USA
Deep learning methods have dramatically improved the ability to detect objects in natural images and semantically segment images for scene understanding. However, these advances have not yet automated the understanding of information contained in hand-drawn figures, technical diagrams, mathematical equations, scanned documents, data plots, and other images conveying technical information (some examples are displayed in Figure 1 below), because those types of images have very different properties from natural images. For this type of image, shape and topology are far more important features than the intensity, color, texture, shadow and shading which dominate natural images.
This workshop will focus on techniques for analysis and retrieval of this type of imagery, which lags far behind the current state of the art for natural images. The design or learning of appropriate representations, such as topological graphs, and of similarity metrics that capture the underlying semantics of such imagery, are of particular interest. Methods for cross-modal retrieval (e.g., retrieval of a photographic image from a hand-drawn sketch) are also relevant.
(Hover the figure to see zoomed-in details.)